get a red build green
CI is red and you need the smallest change that makes it pass — not a refactor that happens to go green.
recs
| model | best when | avoid when | cost vibe |
|---|---|---|---|
| Sonnet-class | The failure is a real bug: flaky test, race, wrong assertion, or the log doesn’t match the diff. | The log already says the exact line and the fix is one character. | $$$ |
| Faster / cheaper cloud | Lint, type errors, a missing import, a snapshot that should update. | “Just make it pass” would mean deleting the test. | $ |
| Local mid | Private repo, and the log is short enough to paste. | The failure only reproduces in CI, not on your machine. | ~free |
why
- A green build is not the job. The job is the failure the test was written to catch.
- Cheap models silence the error. Frontier earns it when the red is a real regression and the tempting fix is to weaken the check.
- Read the log before the diff. Most red builds are one cause, not five.
- Pair with test-generation only after it’s green for the right reason.